After the twenty patterns were serially learned by the network, a test was made of the
percentage of these items were exactly recognized (i.e., actual output within the criterion of
0.2 of the desired output). Figure 4 shows that with standard backpropagation this figure is
around 40% whereas it climbs rapidly to 65% with the addition to each new pattern to be
learned with only 5 pseudopatterns from final storage. With 50 pseudopatterns added to
each new item to be learned, exact recognition of all of patterns climbs to 85%